OWASP Top 10 for AI/LLM Applications (2025)

Risk	Description	Mitigation Strategy
LLM01: Prompt Injection	Manipulating model inputs to alter behavior	Input sanitization, prompt templates, output validation
LLM02: Sensitive Information Disclosure	Model leaks training data or secrets	Data sanitization, output filtering, access controls
LLM03: Supply Chain	Compromised training data/models	Vendor assessment, model verification, provenance tracking
LLM04: Data Poisoning	Malicious training data injection	Data validation, model monitoring, retraining controls
LLM05: Improper Output Handling	Unsafe output processing	Output encoding, validation, sandbox execution
LLM06: Excessive Agency	Over-privileged AI systems	Least privilege, approval workflows, scope limitations
LLM07: System Prompt Leakage	Exposing system instructions	Prompt protection, monitoring, access controls
LLM08: Vector Weaknesses	ML model vulnerabilities	Adversarial testing, input validation, model hardening
LLM09: Misinformation	AI-generated false information	Fact-checking, source attribution, confidence scoring
LLM10: Unbounded Consumption	Resource exhaustion	Rate limiting, cost controls, request validation